Research On Underwater Seafood Detection Method Based On Deep Learning

China is rich in underwater marine product resources,and the development of marine resources has received extensive attention in recent years.In this context,the importance of accurately identifying underwater organisms has become increasingly prominent.Currently,the catch of marine products such as sea cucumbers and scallops mainly depends on manual diving,which poses serious risks to the health of the workers and cannot guarantee their safety.Therefore,the use of intelligent target detection technology to achieve underwater machine fishing marine products has become a new direction of the current marine aquaculture.However,due to the special underwater environment and lighting conditions,interference factors such as impurities and microorganisms exist.In addition,the different living habits of underwater organisms make the collected image targets more dense,and the targets cover each other and are small,which is easy to miss detection and false detection.In response to this issue,this thesis conducted an in-depth analysis and research on existing underwater marine product target detection algorithms,aiming to improve their low detection accuracy and inefficient problems.Two models were built to enhance the detection effect of underwater marine products from two aspects:improving detection accuracy and speed.The main research of this thesis is as follows:To address the issues of densely packed underwater targets,similarity between targets and the background,and small size targets leading to false detection and missed detection,this thesis presents an improved YOLOv5(You only look once5)algorithm to enhance the accuracy of seafood detection.Based on the YOLOv5 target detection algorithm,we constructed two types of C3 modules.One is the C3 module with improved high-level features,which combines the C3 module of high-level features with Swin Transformer,named as C3 STR module.This module expanded the receptive field of the network,enabling a better acquisition of global information and increasing feature extraction capabilities.The other one is the C3 module with improved low-level features.Firstly,the input feature layer of the C3 module with low-level features is divided into two parts according to channels,each part goes through convolution and then integration,so as to enhance the feature extraction ability of the network model while only slightly increasing computation,this improved module called DC3 module.In addition,this thesis also improved the original network’s feature fusion module,merging the features of the PAN(Path Aggregation Network)part with the features of the output detection part,which aids in better integration of high-level and low-level information,further enhancing the detection ability of the network model.Through testing on images from the seafood dataset used in this thesis,the result was m AP(mean Average Precision)of 88.3%,which is 2% higher than the original YOLOv5 algorithm.Compared with other single-stage target detection algorithms SSD(Single shot multi Box detector)and YOLOv4-tiny(You only look once version 4-tiny),the accuracy of the improved algorithm increased by 8.4% and 9.1%,respectively.This not only enhanced detection precision but also reduced missed detections due to severe target blockage.This thesis proposed an improved lightweight YOLOv5 model to address the low real-time performance of underwater target detection models.Firstly,this thesis adopts the lightweight network Mobile Netv3 as the backbone network of YOLOv5.At the same time,due to the SIo U(SCYLLA-Intersection over Union)loss function introducing the vector angle between the real box and the predicted box,using the SIo U loss function can more accurately identify targets.This improved network model is named MS-YOLOv5.Although some detection accuracy was inevitably sacrificed in the process of improving detection speed,overall,the number of model parameters was reduced.This model can effectively balance the speed and accuracy of underwater product detection,achieving the expected research objectives.